Document to Markdown: AI OCR that keeps tables and columns

Busy-hour traffic forecasting for mobile cells

Technical note 7, revision 3 · October 2026

1. Problem

An operator plans capacity per cell from the busy-hour downlink volume. The busy hour is the sixty-minute window of a day with the highest traffic; for a cell it moves with the week and with the season. We forecast the 95th percentile of the daily busy-hour volume over the next 13 weeks from 52 weeks of history, per cell and per band.

2. Features

Only past-window features are used: the trailing 4-, 8- and 13-week medians, the week-of-year, the share of 5G-capable devices, and the PRB utilisation at the busy hour. Labels come from the future window only, and validation is out of time: the last 13 weeks are held out.

$$\hat{y}_{t+h} = \alpha + \Sigma_k \beta_k x_{t,k} + \gamma \cdot s(t+h) + \varepsilon$$

3. Results

| Model | MAPE | p95 err | Lift@10% |
| :--- | :--- | :--- | :--- |
| Seasonal naive | 18.4% | 41.2% | 1.00 |
| Gradient boosting | 11.7% | 26.9% | 2.31 |
| GB + device share | 10.9% | 24.1% | 2.58 |

Ranks and lift are reported rather than raw probabilities because the model is not calibrated. The lift at the top decile is what the planners act on: those cells get the first PRB expansions.

4. Caveats

Column names are not concepts. In one data set the column labelled "DL volume" held an hourly average, in another a daily sum; the values, units and a recomputation from raw counters settle the meaning, never the name. Every workbook has several sheets and all of them were read.

KARAK & CO. CAFE
Al Sadd, Doha · Tel 4444 1234
14/10/2026 08:42 · Table 7 · Srv: Ali

2 x Karak tea 10.00
1 x Zaatar manakish 12.00
1 x Cold brew L 18.00
1 x Date cake 14.00

Subtotal 54.00
Service 10% 5.40
TOTAL QAR 59.40

Paid: VISA **** 4821
Auth 087233 · Thank you!